1. technology and computing

The Challenge of Cloud Control
Maria Kihl
Dept. of Electrical and Information Technology, Lund University, Sweden
Erik Elmroth, Johan Tordsson
Dept. of Computer Science, Umeå University, Sweden
Karl Erik Årzén, Anders Robertsson
Dept of Automatic Control, Lund University, Sweden
Abstract
Today’s cloud data center infrastructures are not even near being able to cope with the enormous and rapidly varying capacity demands that will be reality in a near future. So far, very little is understood about how to transform
today’s data centers (being large, power-hungry facilities, and operated through heroic efforts by numerous administrators) into a self-managed, dynamic, and dependable infrastructure, constantly delivering expected QoS with reasonable operation costs and acceptable carbon footprint for large-scale services with sometimes dramatic variations
in capacity demands. In this paper, we discuss some of the major challenges for resource-optimized cloud data center. We propose a new research area called Cloud Control, which is a control theoretic approach to a range of cloud
management problems, aiming to transform today´s static and energy consuming cloud data centers into selfmanaged, dynamic, and dependable infrastructures, constantly delivering expected quality of service with acceptable
operation costs and carbon footprint for large-scale services with varying capacity demands.
1. Introduction
During recent years, several dramatic events have
shown that the capacity requirements of online services
can increase enormously within minutes. Extreme examples are the CNN web-site during early reporting of
the 9/11 terror attack (load doubled every 7th minute
with peak 2000% over normal) and the 2500% load
increase at Al-Jazeera’s web-site during the fourth day
of the Egyptian revolution, and the massive peak just
minutes after president Mubarak resigned after 18 days.
With cloud data centers currently growing to tens and
hundreds of thousands of interconnected servers and the
increasing complexity of interconnected applications,
system management is growing in complexity to a scale
necessitating new behavioral abstractions and models
for autonomic computing [1]. Due to the nonlinearity of
emergent local behavior, it is intrinsically challenging
to understand the correlations between local and global
behavior and the effects of local and global management actions. Hence, new behavioral and managerial
abstractions are needed, together with methods for distributed or hierarchical control.
By combining virtualization, distributed systems, optimization techniques, accurate performance models and
autonomic computing, our vision is that today’s cloud
technology can be turned into a self-managed optimized
infrastructure providing unprecedented gains in terms
of cost-efficiency, flexibility, robustness, eco-efficiency
and sustainability [2]. In order to facilitate management, virtual machines (VMs) are typically used as a
basic building block, allowing management tools to
migrate, start, stop, and hibernate instances. Depending
on the cloud deployment scenario, the physical infrastructure that hosts the VMs can be internal, external or
a combination of both. In either case, it should appear
to the cloud user as a single system always delivering
sufficient capacity.
2. Challenges
In this paper, we address a range of important interrelated and algorithmically challenging management
and control problems focused on cloud data centers, as
illustrated in Figure 1.
The first challenge is Performance models, which are of
paramount importance for the design and development
of robust control systems. Notably, cloud data centers
are much larger in scale and with a larger variety of
workload dynamics than previous telecom and Internet
systems. Previous results on server systems have shown
that the dynamics may be captured using a black-box
approach and rather simple queue models [3] or using
flow models [4]. However, for clouds, only a few per-
Figure 2 The proposed research area of Cloud Control.
Figure 1 An illustration of a cloud data center.
formance modeling attempts have been presented, see
for example [5].
The second challenge is Service admission control,
which is in essence long term capacity planning where
the control system decides to admit a service or not
based on anticipated profit. For clouds, only little work
has been performed [6].
Further, Elasticity controllers should allocate enough
resources to a running application to provide acceptable
QoS while avoiding costly over-provisioning. Few current solutions actually go beyond simple reactive (nonpredictive) threshold-based allocation adjustment [7].
These solutions fall into three broad categories; solutions based on machine learning algorithms, solutions
based on control theory and solutions based on statistical workload analysis.
Also, there are a number of VM placement problems to
be solved to determine where to deploy new VMs and
which VMs to shut down [8]. Given the dynamic nature
of clouds, with significant changes over time both in
demand and supply, VM placement decisions need to
be renewed frequently. Recent work on re-placement
includes performance modeling of VM migration and
performance steering by restricting the number of VMs
that can be migrated concurrently. Other dynamic approaches include use of stochastic integer programming
to handle uncertainty and genetic algorithms to reduce
re-migrations upon load fluctuations.
Another challenge is that large data centers providing
cloud services come not only with the cost of buying
the equipment but also with a substantial increase in
energy costs, which means that energy optimization will
be crucial [9]. Also, new cyber-physical system categories are emerging where energy interactions between
the computing and physical system components are
non-negligible [10]. It is important to understand, control, and optimize the energy/temperature interactions at
machine room scale (i.e., not only amount of resources
but also physical distribution), since these issues are
important prerequisites to reducing the energy cost and
carbon footprint of cloud systems.
Finally, orchestrating the work of all these different
controllers (“holistic management”) is a difficult challenge that has not received much attention [11]. Given
that workloads and performance models are well understood and that “control knobs” are available in terms of
sophisticated methods for performing elasticity control,
admission control, VM placement, and power management, there is still a challenge to optimize the whole
system behavior. Since the “knobs” may interact in
many ways and different performance policies may
interfere with one another, uncoordinated interference
of multi-knobs can lead to instability or poor performance, calling for a holistic view of resource management [12].
Cloud control
The challenges for resource optimized cloud data centers are immense, and we therefore propose a new research area that we call Cloud Control. The area will
address the whole range of closely related challenges
for the efficient autonomic management of resources in
cloud data centers. Focus is on fundamental modeling,
control methods, and algorithms for performing management tasks. The topics are individually important
but also fit into the same bigger picture, as illustrated in
Figure 2. Each challenge described before can be incorporated in a framework for cloud datacenter resource
optimization and control, and some examples will be
presented below.
First, the service admission controller decides whether
to accept the elastic service that over time may have
largely varying capacity demands. A cloud provider
must determine the optimal number of services to admit
in order to maximize its utility (revenue, utilization,
etc.) without endangering Service Level Agreements
(SLAs) of already provisioned services. Each admitted
service adds a stochastic load to the cloud infrastructure. Admission control schemes are thus needed to
assess the long term effects of new services depending
on their estimated workloads.
Second, the objective of the elasticity control system is
to allocate enough so called Compute Units (CUs) in
order to comply with certain control objectives, targeting the performance expectations of the system. We
proposed to define the term CU in order to have a metric for capacity needs, which is independent of the definition of VMs. CUs will later be mapped to suitable
VMs. If it is assumed that the system state and workload can be estimated with some accuracy, the controller design basically becomes a stochastic optimal control problem.
Third, the VM placement controller both perform the
mapping from CUs to an appropriate set of VMs to run
locally or in remote data centers, and determines the
packing of VMs on physical machines. We consider the
mappings from CU to VM and from VMs to datacenters before performing the assignment of VMs to PMs
within the datacenter. These capacity allocation (assignment) problems that can be formulated, e.g., as
constrained linear problems for optimizing cost (monetary or other) subject to a capacity and possibly other
constraints, such that a specific load balancing (dispersion) objective or requirements on affinity and antiaffinity (e.g., in the same datacenter (or host), not in the
same datacenter).
Fourth, the data center energy optimizer models and
optimizes the energy-temperature interactions at machine room scale for VM placement decisions. The energy optimization will require a cyber-physical perspective.
Finally, the holistic management system monitors the
complete system behavior and optimizes the management tools’ concerted operation with respect to the
overall data center management objectives.
Conclusions
Cloud computing has received much attention the last
years, and is envisioned to be used for a wide range of
services. However, the large amount of data centers
required for these cloud services will have immense
challenges of resource management and control. In this
paper, we have presented some of these challenges.
Also we have proposed a new research area of Cloud
Control, which will address the whole range of closely
related challenges for the efficient autonomic management of resources in cloud data centers.
References
[1] J. Kephart and D. Chess. The vision of autonomic
computing. Computer, 36(1):41–50, 2003.
[2] M. Armbrust et al. A view of cloud computing.
Communications of the ACM, 53(4):50–58, 2010.
[3] J. Cao, M. Andersson, C. Nyberg, and M. Kihl.
Web server performance modeling using an
M/G/1/K*PS queue. In 10th International Conference on Telecommunications, 2003.
[4] D. Tipper and M. K. Sundareshan. Numerical
methods for modeling computer networks under
nonstationary conditions. IEEE Journal on Selected
Areas in Communications, 8(9), 1990.
[5] H. Khazaei, J. Misic, and V. Misic. Performance
analysis of cloud computing centers using
M/G/m/m+r queueing systems. IEEE Transactions
on Parallel and Distributed systems, Vol. 23, No. 5,
May 2012.
[6] K. Konstanteli, T. Varvarigou, and T. Cucinotta.
Probabilistic admission control for elastic cloud
computing. In IEEE Int. Conf. on Service-Oriented
Computing and Applications (SOCA), 2011, 2011.
[7] W. Iqbala, M. Daileya, D. Carrerab, and P. Janeceka. Adaptive resource provisioning for read intensive multi-tier applications in the cloud. Future
Generation Computer Systems, Vol. 27, No. 6,
2010.
[8] J. Tordsson, R. S. Montero, R.M. Vozmediano, and
I.M. Llorente, Cloud brokering mechanisms for optimized placement of virtual machines across multiple providers, Future Generation Computer Systems, 28(2), 358-367, 2012.
[9] A. Greenberg, J. Hamilton, D. Maltz, and P. Patel.
The cost of a cloud: Research problems in data
center networks. ACM SIGCOMM Computer
Communication Review, 39(1), 2009.
[10] M. Poess and R. Nambiar. Energy cost, the key
challenge of today’s data centers: A power consumption analysis of TPC-C results. Proceedings of
the VLDB Endowment, 1(2), 2008.
[11] E. Elmroth et al. . Self-Management Challenges for
Multi-Cloud Architectures, Towards a ServiceBased Internet, Lecture Notes in Computer Science, Vol. 6994, Springer-Verlag, 2011.
[12] J. Heo, P. Jayachandran, I. Shin, D. Wang, T. Abdelzaher, and X. Liu. OptiTuner: On performance
composition and server farm energy minimization
application. IEEE Transactions on Parallel and Distributed Systems, Vol. 22, No. 11, 2011.